Thursday, June 20, 2024

A Push to Address the Pressing Evolutionary Needs of Today’s Data Centers

Marvell Technology, Inc., a leader in data infrastructure semiconductor solutions, has officially announced the launch of its new Alaska® P PCIe retimer product line, which is designed to scale data center compute fabrics inside accelerated servers, general-purpose servers, CXL systems, and disaggregated infrastructure. To kick things off, Marvel will bring two products in 8- and 16-lane PCIe Gen 6 retimers, both capable of connecting AI accelerators, GPUs, CPUs, and other components inside server systems. Before digging any further into this development, though, we first need to acknowledge how AI and machine learning applications are driving data flows and connections inside server systems at significantly higher bandwidth. Such a piece of reality is, in turn, mandating PCIe retimers to meet the required connection distances at faster speeds. Making matters even more complicated is the fact that these AI models are doubling their computation requirements every six months, and therefore, they are now the primary driver of the PCIe roadmap, with PCIe Gen 6 becoming the centerpiece requirement. More on the PCIe Gen 6 would reveal how it operates at 64 gigatransfers-per-second (GT/s), becoming the first ever PCIe standard to use four-level pulse-amplitude modulation (PAM4) signaling, displacing the non-return to zero (NRZ) modulation in the last 20 years. Having said so, higher bandwidth and faster data rate also limits the physical reach that signals can travel reliably, thus reducing the distance connections can span.

Enter Marvell® Alaska P retimers. These retimers compensate for the signal degradations and regenerate the same to deliver reliable communication over physical distances. The fact in question is applicable for connections between GPUs and CPUs within an AI server, between GPUs on different boards, or between CPUs, and a pool of shared memory enabled by CXL. On top of that, the stated retimers can also be used on AI accelerator baseboards, server motherboards, riser cards, or integrated into active electrical cables (PCIe AEC) and active optical cables (PCIe AOC) for emerging multi-rack server system architectures.

“Marvell is the industry leader for data center-to-data center, switch-to-switch, rack-to-rack, server-to-switch, and server-to-server connectivity. We are now entering the market for compute fabrics as PCIe and CXL go through an inflection point, migrating from NRZ to PAM4 technology,” said Venu Balasubramonian, vice president of product marketing, Connectivity Business Unit at Marvell Technology. “Marvell is building on more than 10 years of in-house expertise in PAM4 technology and our industry-leading 5nm PAM4 IP portfolio to enable this transition.”

In terms of more specific features, Marvell’s retimers are understood to be compatible with PCI Express® Gen 6/5/4/3/2/1 and Compute Express Link™ 3/2/1.1. They further boast industry-leading PAM4 SerDes performance, and at the same time, house a low-latency mode for cache-coherent links. Markedly enough, these retimers also have the industry’s lowest power consumption (10W PCIe 6 x16). Apart from that, they deliver at your disposal capabilities like advanced telemetry and diagnostics: in-band FEC monitoring, out-of-band SerDes eye monitoring, embedded logic analyzer, and software suite for fleet management in large-scale deployments.

“Data center operators need new solutions to support evolving workloads and architectures. For instance, customers are looking for PCIe optical cables for longer reach large-scale AI accelerator interconnect and CXL disaggregation infrastructure where electrical connectivity reaches its limits. InnoLight and Marvell are leveraging our strong optical connectivity expertise and collaboration history to move the optical PCIe ecosystem forward,” said Osa Mok, chief marketing officer at InnoLight